A method for clustering metagenomic sequences using a two-level probabilistic model

Abstract

The present invention relates to the technical field of metagenomic sequence clustering, and in particular to a method for clustering metagenomic sequences using a double-layer probability model, including P1, clustering all sequences in an initial metagenomic sequence using a first-layer probability model to obtain multiple primary clusters; P2, clustering each primary cluster again using a second-layer probability model to obtain a final cluster; the second-layer model includes a seed selection model, a k-mer frequency probability model, and a coverage probability model. The method for clustering metagenomic sequences using a double-layer probability model adopted by the present invention uses a two-layer model to process all sequences in the initial metagenomic sequence, can effectively utilize the features of different dimensions of metagenomic sequences, and is applicable to all metagenomic sequencing data, such as intestinal microbial data, soil microbial data, and water microbial data.

Description

Technical Field

[0001] The present invention relates to the field of metagenomic sequence clustering technology, and specifically to a method for clustering metagenomic sequences using a double-layer probability model. Background Art

[0002] As the oldest existing life form on Earth, prokaryotes play an important role in the process of biological evolution, material circulation and environmental change on Earth. Mapping complete bacterial and archaeal genomes is the basis for studying their systematic classification, and is also key data for studying microbial evolution, ecology, and community structure and function. With the development of sequencing technology and metagenomic assembly technology, it is possible to recover microbial sequences from sequencing data, but due to the complexity of real environmental samples, the metagenomic sequences obtained by sequence assembly are highly fragmented. Matching the sequences in the metagenome with the existing sequences in the database to obtain sequence tags (for example, microbial systematic classification names) is an effective way, but since the discovery of microorganisms, only 2.1% of the genomes have been sequenced, and the proportion of complete genome sequences in the database is even smaller. In order to discover new microorganisms, it is urgent to propose practical new methods and means to deeply and systematically mine microbial sequencing data.

[0003] Using clustering ideas to cluster fragmented sequences to the genome level is the most ideal approach at this stage. The information (features) used for clustering mainly include the nucleotide composition of the sequence ( -mer frequency) and sequencing data coverage. -mer frequency information is species-specific, and different regions belonging to the same genome have similar -mer frequency information, and the relative abundance of microorganisms in the sample can also be used to determine the origin of the sequence. In addition, certain molecular markers, such as genes with a single copy number in the microbial genome, are also used for clustering.

[0004] Existing clustering methods do not use the above three features reasonably and effectively, which is manifested in unreasonable weight settings for different features, the inability of probability models to fit features well, and imperfect method construction. Although existing methods have achieved good results on simulated data and real samples, there is still a big gap from the ideal results, mainly manifested in the low integrity and high contamination of the genomes generated by clustering. Therefore, it is necessary to develop new clustering algorithms for microbial sequences. Summary of the invention

[0005] In view of the problems existing in the prior art, the present invention provides a method for clustering metagenomic sequences using a double-layer probability model.

[0006] The present invention is achieved through the following technical solutions:

[0007] A method for clustering metagenomic sequences using a two-layer probability model comprises the following steps:

[0008] P1, clustering all sequences in the initial metagenome using the first-layer probability model to obtain multiple primary clusters; the first-layer probability model is a DPGMM model;

[0009] P2, using the second-layer probability model to cluster each primary cluster again to obtain the final cluster; the second-layer probability model includes a seed selection model, -mer frequency probability model and coverage probability model.

[0010] Preferably, in P1, the following steps are included:

[0011] P11, obtained from the initial metagenome -mer frequency feature vector and coverage feature vector, where -mer means the length is oligonucleotides;

[0012] P12, -mer frequency feature vector and coverage feature vector are merged into a single feature vector;

[0013] P13, using the Dirichlet process to build the DPGMM model, using the merged feature vectors from P12 as input for clustering;

[0014] P14, using the variational method to iteratively estimate the parameters of the DPGMM model constructed in P13, and using the estimated DPGMM model parameters to obtain primary clustering.

[0015] Preferably, in P11, the -mer frequency feature vector The value of is 4.

[0016] Preferably, in P14, during the iteration process, the weights of all clusters are calculated every 10 cycles; then all clusters are divided into large clusters and small clusters according to the size of the weights, wherein the weight of the large cluster is greater than or equal to 500 Kb, and the weight of the small cluster is less than 500 Kb; then the small clusters are disbanded and reallocated to the large clusters; finally, the reassembled large clusters are re-iterated.

[0017] Preferably, in P14, the weight of the primary cluster is greater than or equal to 500 Kb.

[0018] Preferably, in P2, the following steps are included:

[0019] P21, using the seed selection model to select from each primary cluster A seed sequence, is a positive integer;

[0020] P22, the seed sequence is used The classification probability obtained by the -mer frequency probability model and the coverage characteristics obtained by the coverage probability model are combined to obtain the final clustering.

[0021] Preferably, in P22, the steps of obtaining the classification probability are as follows: first, using the multi-classification support vector model seed sequence -mer frequency features are used for training to obtain a trained classification model; the trained classification model is then used to predict all sequences in the primary cluster and obtain the corresponding classification probabilities.

[0022] Preferably, in P22, the final clustering is obtained using an expectation-maximization algorithm.

[0023] Preferably, the final clustering is benchmarked using a simulated data set and a real data set to obtain a clustering result.

[0024] Preferably, the clustering results are evaluated using a precision-recall evaluation system.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] The method of clustering metagenomic sequences using a two-layer probability model in the present invention uses a two-layer model to process all sequences in the initial metagenome, which can effectively utilize the characteristics of different dimensions of metagenome sequences and is applicable to all metagenome sequencing data, such as intestinal microbial data, soil microbial data, and water microbial data.

[0027] Furthermore, during the variational iteration process, clusters smaller than 500 Kb are dissolved and redistributed in order to reasonably balance the feature weights and help improve the clustering quality.

[0028] Furthermore, the classification performance of the model was evaluated using the precision-recall evaluation system. The precision was expressed as the ratio of the total length of the correct clustering sequences to the cluster size, reflecting the size of the false positive clustering results (the degree of cluster contamination); the recall was expressed as the ratio of the total length of the correct clustering sequences to the corresponding genome size, which was used to describe the completeness of the program's predicted clustering.

[0029] The model reported in the present invention has been fully verified on test data sets and real data sets. Compared with other existing methods, the present invention can produce more high-quality clusters, providing a prerequisite for subsequent metagenomic data analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 The present invention is a computational flow chart of a method for clustering metagenomic sequences using a double-layer probability model. DETAILED DESCRIPTION

[0031] The present invention is further described in detail below in conjunction with specific embodiments, which are intended to explain the present invention rather than to limit it.

[0032] A method for clustering metagenomic sequences using a two-layer probability model comprises the following steps:

[0033] P1, cluster all sequences in the initial metagenome using the first-level probability model (DPGMM model) to obtain multiple primary clusters.

[0034] P11, obtained from the initial metagenome -mer frequency feature vector and coverage feature vector, where -mer means the length is The oligonucleotides -mer frequency feature vector The value of is 4;

[0035] P12, -mer frequency feature vector and coverage feature vector are merged into a single feature vector;

[0036] P13, using the Dirichlet process to build the DPGMM model, using the merged feature vectors from P12 as input for clustering;

[0037] P14, using the variational method to iteratively estimate the parameters of the DPGMM model constructed in P13, and then using the estimated DPGMM model parameters to obtain primary clusters with weights greater than or equal to 500 Kb.

[0038] The specific model of the DPGMM model is as follows:

[0039] in is the input feature matrix, is the parameter set, denote the total number of sequences and genomes, respectively. The parameter is expressed as of distributed, is the scale parameter of DP, is a random variable. The parameter is expressed as Normal-inverse-Wishart distribution, Genome The mean and covariance of is represented as the prior mean, is expressed as the cluster mean ( ) distribution, Expressed as The prior, Expressed as a prior on the degrees of freedom of the Wishart distribution. The parameter is expressed as The classification distribution of Represented as genome The weight of Represented as a sequence Source. sequence exist The probability density of a Gaussian distribution.

[0040] During the iteration process, the weights of all clusters are calculated every 10 cycles; then all clusters are divided into large clusters and small clusters according to the size of the weights, where the weight of the large cluster is greater than or equal to 500Kb and the weight of the small cluster is less than 500Kb; then the small clusters are disbanded and reallocated to the large clusters; finally, the reassembled large clusters are re-iterated.

[0041] The variational distribution is as follows:

[0042]

[0043] in, is the parameter to be estimated, where Also represented as a sequence Belong to the genome probability.

[0044] P2, using the second-layer probability model to cluster each primary cluster again to obtain the final cluster; the second-layer probability model includes seed selection model, -mer frequency probability model and coverage probability model. Specifically include the following steps:

[0045] P21, using the seed selection model to select from each primary cluster A seed sequence, is a positive integer;

[0046] P22, the seed sequence is used The classification probability obtained by the -mer frequency probability model and the coverage characteristics obtained by the coverage probability model are combined, and the final clustering is obtained using the expectation maximization algorithm.

[0047] In P22, the steps to obtain the classification probability are as follows: First, use the multi-classification support vector model seed sequence -mer frequency features are used for training to obtain a trained classification model; the trained classification model is then used to predict all sequences in the primary cluster and obtain the corresponding classification probabilities.

[0048] The likelihood function of the model to obtain the final clustering is as follows:

[0049]

[0050] in is the input feature matrix, is the parameter set, denote the total number of sequences and genomes, respectively. Represented as a sequence In P22, we get the classification probability. Represented as a sequence of -mer frequency probability and sequence Specific a priori cluster weights of distance, A positive number used to balance -The weight of mer frequency feature and coverage feature.

[0051] P3, after obtaining the final clustering, the final clustering is benchmarked using simulated data sets and real data sets to obtain the clustering results, and the clustering results are evaluated using the precision-recall evaluation system.

[0052] Taking the high-complexity benchmark dataset provided by CAMI as an example, refer to Figure 1 , using the method of the present invention, metagenomic sequence clustering is performed.

[0053] like Figure 1 As shown, the present invention provides a two-layer probability model for metagenomic sequence clustering, comprising the following steps P1 - P3.

[0054] P1: Use the modified DPGMM model to cluster all initial sequences to generate primary clusters.

[0055] P2: For each primary cluster generated by P1, use the second-level probability model 9 (seed selection model, -mer frequency probability model and coverage probability model) to make the final clustering of the sequences.

[0056] P3: Cluster the CAMI high complexity benchmark dataset based on the final clustering and perform performance evaluation on the results.

[0057] Specifically, it includes: using the benchmark evaluation system provided by CAMI to evaluate all clusters generated by the present invention, counting the number of genomes identified under different threshold conditions, and comparing with existing methods.

[0058] Experimental results: For the CAMI high-complexity benchmark dataset, this method identified 272 clusters with precision higher than 0.90 and recall higher than 0.50, including 190 high-quality clusters ( ), and the best results among the existing methods are 235 and 168.

[0059] The model reported in the present invention has been fully verified on test data sets and real data sets, and performs well in benchmark tests. Compared with other existing methods, the present invention can produce more high-quality clusters, providing a prerequisite for subsequent metagenomic data analysis.

Claims

1. A method for clustering metagenomic sequences using a two-layer probability model, characterized in that: The following steps are involved: P1, clustering all sequences in the initial metagenome using the first-layer probability model to obtain multiple primary clusters; the first-layer probability model is a DPGMM model; Specifically: P11, all sequences were obtained from the initial metagenome -mer frequency feature vector and coverage feature vector, where -mer means the length is oligonucleotide; -mer frequency feature vector The value of is 4; P12, -mer frequency feature vector and coverage feature vector are merged into a single feature vector; P13, using the Dirichlet process to build the DPGMM model, using the merged feature vectors from P12 as input for clustering; P14, using the variational method to iteratively estimate the parameters of the DPGMM model constructed in P13, and using the estimated DPGMM model parameters to obtain primary clustering; in the iterative process, the weights of all clusters are calculated every 10 cycles; then all clusters are divided into large clusters and small clusters according to the size of the weights, where the weight of the large cluster is greater than or equal to 500 Kb, and the weight of the small cluster is less than 500 Kb; then the small clusters are disbanded and reallocated to the large clusters; finally, the reassembled large clusters are re-iterated; the weight of the primary cluster is greater than or equal to 500 Kb; P2, using the second-layer probability model to cluster each primary cluster again to obtain the final cluster; the second-layer probability model includes a seed selection model, -mer frequency probability model and coverage probability model; Specifically: P21, using the seed selection model to select from each primary cluster A seed sequence, is a positive integer; P22, the seed sequence is used The classification probability obtained by the -mer frequency probability model is combined with the coverage characteristics obtained by the coverage probability model to obtain the final clustering; The steps for obtaining the classification probability are as follows: first, the multi-classification support vector model seed sequence is used -mer frequency features are used for training to obtain a trained classification model; the trained classification model is then used to predict all sequences in the primary clustering and obtain the corresponding classification probabilities; the expectation maximization algorithm is used to obtain the final clustering.

2. The method for clustering metagenomic sequences using a two-layer probability model according to claim 1, characterized in that: After obtaining the final clustering, the final clustering is benchmarked using simulated and real datasets to obtain the clustering results.

3. The method for clustering metagenomic sequences using a two-layer probability model according to claim 2, characterized in that: The clustering results are evaluated using the precision-recall evaluation system.

Images